Interpreting and Steering the Polar Vortex in a Weather Foundation Model

Machine learning weather prediction (MLWP) models can predict the atmospheric state efficiently but are still "black boxes". We investigate whether large-scale atmospheric phenomena such as the polar vortex are (1) represented internally and (2) can be manipulated within the Aurora foundation model. By adapting contrastive activation addition to Aurora's latent space, we show that latent injection can intensify or weaken the polar vortex. We also quantify the physical consistency of these steered forecasts.